Physicists Split on AI Use in Peer Review
The use of generative artificial intelligence (AI) is permeating all domains of science—from analyzing experimental data to writing manuscripts. And it is now quickly spreading to the peer-review system used to scrutinize papers for publication. But does this use concern researchers in the physical sciences? A recent analysis from a major physics publisher aimed to find out.
The Institute of Physics Publishing (IOPP) in the UK had carried out a similar survey in 2024 involving more than 3000 participants. But the rapid spread of AI and the evolution of publisher policies motivated a follow-up study. This year’s survey found that 41% out of the 348 respondents—who had either published in an IOPP journal or refereed for one within the six months prior—thought generative AI had an overall positive impact on peer review. That figure represents an 11% increase on the IOPP’s 2024 analysis. Meanwhile, 37% of survey respondents thought generative AI was impacting peer review negatively. Just 22% said they were unsure—a decrease from 26% since the 2024 survey.
The detailed responses of survey participants shed light on the reasons behind their attitude. “I fear that many reviewers will use it as a replacement to thinking,” wrote one respondent from the negative camp. Other respondents, however, stressed some advantages that generative AI may offer. “It will make the process faster. It will make it easier to understand the key concepts of an article, saving valuable time and allowing for better focusing on what is novel and [what must] be checked thoroughly,” wrote another surveyed researcher.
“We see a polarization within the physical sciences community about the use of AI in peer review,” says Laura Feetham-Walker, IOPP’s reviewer-engagement manager. “Given that we basically asked roughly the same population of people the exact same question [a] year apart, I think what we can say is more people are informed, and more people have an opinion,” she says.
Brian Earp, a philosopher and bioethicist at the National University of Singapore who has written about AI and the future of peer review, agrees on the origin of the increasing polarization of opinions. He adds that it’s only natural that AI is making inroads, as it offers ways to alleviate some of the burden of doing a peer review, which is usually not compensated.
According to the survey, 57% of researchers said that they would be unhappy if referees used generative AI to conduct peer review on a manuscript they had authored, and 42% thought that they could confidently detect AI-written reports. “The ‘peer’ in ‘peer review’ is really important to a lot of people,” Feetham-Walker notes. “Researchers, on the whole, actually want their papers to be reviewed by human experts, and that matters to them.”
Although many researchers would not like referees of their papers to use AI, about a third of surveyed IOPP authors admitted to using AI tools as they acted themselves as referees. This occurs despite the fact that IOPP currently prohibits referees from using generative AI. But with much of the publishing industry now allowing researchers to use AI to translate or to polish writing, IOPP is taking a second look at its own policy.
“We are looking at it based on a pragmatic perspective and also [based on] the understanding that there’s a potential to make people’s lives a bit easier,” Feetham-Walker says. But she says that IOPP intends to design a policy that doesn’t upset or alienate authors. Feetham-Walker also notes that AI-generated reports are not of good quality and lack depth and expertise. “Large language models are just knitting words together,” she says. “They’re not really doing much logical reasoning at all.”
IOPP staff can easily work out which referee reports were authored entirely using generative AI, Feetham-Walker says. “We can spot them,” She adds that they also routinely ask ChatGPT to review a manuscript, just to check in and see how fast these tools are developing.
Earp, for one, has used AI to help him carry out peer review. To do so, he put his handwritten notes on manuscripts he is reviewing into ChatGPT and asked the tool to draft a summary section, which he then edited so that it said precisely what he wanted it to say. “It seems like a reasonable use of AI that benefits me and doesn’t harm the author,” he adds. Earp notes, however, that he has spotted reviews on his own papers that were written entirely by AI. Such reviews weren’t helpful and probably hadn’t been vetted properly by journal editors, he says.
Daniel Ucko, Head of Ethics and Research Integrity at the American Physical Society (APS), says that the APS does not allow authors to upload manuscripts onto AI chatbots. (APS is the publisher of Physics Magazine.) “There are confidentiality issues about that, and we don’t really know where the information is going,” he notes. What’s more, authors haven’t given permission for their manuscript to be transmitted in that way, he says.
Since the launch of AI chatbots, a number of lawsuits against AI companies have surfaced, claiming that the firms illegally accessed research papers, among other content, to train AI models. Last month, the US-based AI-company Anthropic reached a first-of-its kind settlement with copyright holders, agreeing to pay $1.5 billion in damages.
“The meaning of academic publishing is to have researchers talk to other researchers, and to enable that,” Ucko says. “If we just have AI models talking to each other, we’re wasting everybody’s time.”
–Dalmeet Singh Chawla
Dalmeet Singh Chawla is a freelance science journalist based in London, UK.